AI Capital Expenditure, Custom ASICs, and Domestic Chinese Computing Power Jointly Support a Strong Semiconductor Outlook

Overview

Starting with global cloud capital expenditure and aggregate AI chip demand, the report sequentially analyzes GPUs and custom ASICs, TSMC's advanced nodes and CoWoS, HBM and conventional memory, domestic AI computing power in China, and post-quantum cryptography migration. The core conclusion is that AI remains the semiconductor industry's strongest growth engine, but non-AI chips, cost inflation, and computing infrastructure bottlenecks are causing pronounced divergence within the industry.

Core views

The report first uses cloud capital expenditure to validate the sustainability of AI semiconductor demand. The global semiconductor market is expected to reach US$1.5tn by 2030, with AI semiconductors contributing about half; the AI semiconductor TAM is expected to be approximately US$753bn by 2030, while the supply-chain-data-driven bull-case scenario for the 2026 cloud AI semiconductor market is US$485bn. Capital expenditure by the four largest CSPs—Amazon, Google, Microsoft, and Meta—grew 87% YoY in the second quarter of 2026, while the capital expenditure-to-EBITDA ratio has exceeded 70%. Tracking of the world's 14 largest listed CSPs indicates that cloud capital expenditure will approach US$1.4tn in 2027, before including sovereign AI. The report believes this level of investment will continue to drive demand for GPUs, ASICs, server CPUs, advanced packaging, HBM, and testing, although budget constraints are also becoming more significant. Within the AI infrastructure supply chain, the report expects the value of AI compute wafer consumption to exceed US$46bn in 2027, with NVIDIA still accounting for the majority; HBM consumption could reach as much as 48bn Gb during the same period, with NVIDIA remaining the largest source of demand. TSMC is estimated to have produced 5.1mn related chips in 2025, while full-year GB200 NVL72 rack shipments are expected to reach 30k. Based on announced power deployments, per-rack power consumption, and chip counts, the report derives demand for CoWoS and 2/3nm wafers and believes that persistently strong AI demand could drive TSMC to expand CoWoS capacity to 200kwpm in 2027. The transmission of supply and demand implies joint benefits for advanced nodes, advanced packaging, HBM, I/O, and testing, while supporting supplies such as ABF substrates could constrain the volume ramp of certain AI chips. TSMC is the principal beneficiary of this growth thesis. The report notes that demand from TSMC's N2 customers remains strong, and the company's stated capacity roadmap implies an approximately 70% CAGR in N2 capacity from 2026 to 2028. N5 capacity is expected to decline in 2027, while N3 capacity continues to increase, consistent with the report's demand forecasts. Mature and specialty-node capacity is expected to maintain a 7% CAGR from 2024 to 2029. AI semiconductor revenue is expected to account for more than 30% of TSMC's revenue in 2026 and improve both the revenue mix and margins. In advanced packaging, TSMC CoWoS can support up to approximately 9.5 times reticle size or four chips per wafer; if supply-chain execution proceeds smoothly, Intel EMIB can more readily support large chips exceeding 12 times reticle size. Therefore, scalability and mass-production execution for ultra-large packages are key areas of competition between the two technologies. Beyond general-purpose GPUs, the report emphasizes that CSPs will continue to require custom chips. Different cloud platforms need optimization around training, inference, cost, power consumption, and their own software stacks, so even the powerful performance of NVIDIA GPUs will not eliminate ASIC demand. AWS Trainium forecasts indicate that Alchip's recognizable Trainium revenue could rise from US$1.8bn in 2026 to US$2.8bn in 2027 and reach US$8bn in 2028, driven by Trainium4; by then, Trainium revenue would account for approximately 82% of Alchip's projected US$9.8bn total revenue. For Google TPU, the report forecasts total shipments increasing from 3.7mn units in 2026 to 7.35mn units in 2027 and remaining above 6.5mn units in 2028. Even under a customer-owned-tool model, the report still believes MediaTek could at least be responsible for the packaging and I/O die of TPU v10, with TPU revenue estimated at US$13.5bn in 2027, US$43.5bn in 2028, and US$70bn in 2029. KYEC should benefit from the combined growth in testing volumes for AI GPUs, TPUs, and CPUs. CPUs are also emerging as a new growth segment. NVIDIA Vera uses a CPU core design that is not split across multiple chiplets, enabling faster inter-core connectivity, and the report estimates that its performance could reach 1.8 times that of the highest-performance x86 CPU. In Morgan Stanley's top-down model, the Agentic CPU TAM is expected to expand at a 251% CAGR from F26 to F30, with the bull-case scenario implying a US$238bn CPU orchestration market. This view is based on agentic workloads requiring stronger CPU scheduling, data preparation, and system orchestration capabilities rather than merely increasing GPU counts. The conventional semiconductor cycle is not synchronized with the AI upcycle. The report expects logic foundry utilization to reach 80% in the second half of 2026 and notes that historically, declining days of inventory have typically coincided with increases in semiconductor stock indices. However, excluding memory and NVIDIA AI GPU revenue, non-AI semiconductor growth is expected to decline in 2026. Rising wafer, assembly and testing, and memory costs are creating “chip inflation”; price elasticity could suppress demand for technology products and pressure the margins of chip design companies in 2026. At the same time, the supply chain's prioritization of AI semiconductors is crowding out resources such as T-Glass and memory for non-AI products. In memory, AI storage demand is expected to create NAND shortages, while the shortage of NOR Flash could persist through 2026. The DDR4 shortage is expected to continue into the second half of 2026, although upside in spot prices is constrained. The report assesses supply gaps using wafer spot prices, module prices, and quarterly supply-demand breakdowns, combining capacity demand from AI servers with supplier exits or capacity shifts. The implication is that the memory recovery will not be uniform: the duration of shortages across different categories, the spread between spot and contract prices, and the pace at which suppliers add capacity will continue to determine the degree of earnings realization. China's AI computing power constitutes another independent growth theme. DeepSeek has demonstrated lower-cost inference capabilities, which the report believes will stimulate inference demand; the surge in token usage on ByteDance's Volcano Engine/Doubao is also viewed as evidence of robust demand. The capabilities of domestic foundries and AI GPU supply chains are improving, and the report expects China's AI chip TAM to grow to US$91bn by 2030. A comparison of large-model inference economics shows that domestic Chinese chips have a lower total cost of ownership and can achieve a per-token cost similar to NVIDIA processors; their substantially lower selling prices deliver stronger cost performance. Infrastructure-level clustering is also narrowing the market-perceived technology gap: Huawei's Atlas 950 SuperPod, showcased at WAIC in 2026, can scale to 1,024 Ascend 950DT NPUs and uses hybrid copper and optical interconnects with up to 256TB of pooled memory. The report also expects China's CPU TAM to reach US$42bn by 2030, with Hygon's share of China's server CPU market reaching 18% by 2028. The quantum security section focuses on migration from existing asymmetric cryptography to post-quantum cryptography. The report notes that governments have established PQC migration roadmaps, with major economies broadly aligning with NIST standards, although progress varies by region; 2030 is a key inflection point. Because asymmetric encryption is widely used in modern electronic systems, migration will affect everything from chips and devices to various connectivity layers. Hardware security is foundational to the quantum supply chain, and PUF can provide devices with a hardware root of trust. The report estimates that the PUF TAM will reach approximately US$310mn by 2030 and assesses the valuation of related assets using 2027 P/E ratios for global semiconductor and PUF IP-related companies and eMemory's historical P/E range. The report's overall industry positioning is positive, but its internal selections are clearly differentiated. MediaTek is identified as the Top Pick for AI, with other key AI names including TSMC, SMIC, Aspeed, Alchip, KYEC, ASE, FOCI, ASMPT, AllRing, and GUC. AP Memory is the Top Pick for memory, with Macronix, Nanya Tech, Winbond, and GigaDevice also highlighted. The China AI/semiconductor/equipment list includes Iluvatar, Cambricon, Hygon, NAURA Tech, and AMEC; the testing equipment and consumables list includes WinWay, MPI, Hon Hai Precision, and Gudeng; and the mature-node list includes UMC. Hygon is also explicitly rated OW; OmniVision, Phison, MetaX, Realtek, and GlobalWafers are rated EW, while WIN Semi, Silergy, and ASMedia are rated UW.

Analysis framework

The report first estimates AI computing demand using CSP capital expenditure, power deployments, and rack counts, then breaks that demand down into GPU/ASIC quantities, wafers, advanced nodes, CoWoS, HBM, and testing volumes. It subsequently forecasts the revenue exposure of companies such as Alchip, MediaTek, and TSMC based on company projects and customer-generation roadmaps. For China, it assesses domestic substitution by combining token demand, chip prices, total cost of ownership, per-token cost, and SuperPod system capabilities. The quantum security section estimates the PUF market based on regional PQC migration timetables, connectivity layers, and demand for hardware roots of trust, and compares related assets using P/E ratios.

Methodology notes

• Industry/Sector Analysis FrameworkSupply-demand framework

  Supply-demand models for AI chips, CoWoS, HBM, NAND, NOR, and DDR4

  The report converts end-market computing demand into chip, wafer, packaging, and memory consumption, then compares it with capacity and supply growth at each stage to determine shortages, utilization, and price direction.

• Industry/Sector Analysis FrameworkUpstream-Midstream-Downstream Industry Chain Transmission

  Transmission of cloud capital expenditure through the semiconductor supply chain

  Starting with CSP capital expenditure and data-center power deployments, the report sequentially derives demand for racks, GPUs and ASICs, advanced-node wafers, CoWoS, HBM, I/O, and testing.

• Industry/Sector Analysis FrameworkVolume-price decomposition

  Decomposition of ASIC project shipment volumes and per-chip ASP revenue

  Revenue forecasts for Alchip and MediaTek are formed by multiplying shipment volumes for each Trainium or TPU generation by recognizable per-chip ASP, thereby illustrating the contribution of customer projects to company revenue.

• Valuation MethodPE/PEG valuation

  Cross-sectional P/E comparison and historical ranges

  The report compares valuations across foundries, assembly and testing companies, memory companies, IDMs, equipment companies, and fabless designers, and assesses quantum-security-related assets using 2027 P/E ratios and eMemory's historical P/E range.

• Cycle and Industry Conditions FrameworkInventory cycle (Kitchin)

  Relationship between days of inventory and industry conditions

  The report tracks supply-chain days of inventory in the first quarter of 2026 and notes that historically, declines in days of inventory have typically coincided with increases in semiconductor stock indices.

• Competition and Strategy FrameworkValue chain analysis

  Value chain spanning GPUs, ASICs, foundries, packaging, memory, testing, and system integration

  The report identifies each company's responsibilities in AI chip design, I/O dies, wafer manufacturing, advanced packaging, HBM, testing, and SuperPod systems to determine revenue attribution and benefiting segments.

• (Method Outside Vocabulary)

  Comparison of total cost of ownership and per-token cost

  The report combines chip prices, system investment, and inference output to compare the inference economics of domestic Chinese chips with NVIDIA processors rather than comparing only peak performance.

• (Method Outside Vocabulary)

  Deriving chip and wafer demand from power deployments

  The report uses announced power capacity, rack power consumption, project life cycles, and chip counts per rack to derive GPU/ASIC quantities and demand for CoWoS and 2/3nm wafers.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• MediaTek

  Identified as the Top Pick for AI, with the report expecting its TPU revenue to reach US$70bn in 2029.

  Strengths

  Expected to undertake at least the packaging and I/O die for TPU v10 and participate in multiple subsequent TPU generations.

  Weaknesses

  Revenue realization depends on the scale of Google's TPU projects and the recognizable share of value under the customer-owned-tool model.

  Comparison

  Unlike the pure-GPU route, MediaTek's opportunity comes from CSP custom ASICs and advanced packaging value.

  Risks

  Project shipments, mass-production execution, or recognizable ASP could fall below the report's assumptions.

• TSMC

  The core manufacturing provider for AI GPU, ASIC, advanced-node, and CoWoS demand, and included among the key AI names.

  Strengths

  N2 demand is strong, while N3 capacity continues to increase; AI semiconductors are expected to account for more than 30% of 2026 revenue.

  Weaknesses

  N5 capacity is expected to decline in 2027, while capacity expansion must keep pace with rapidly growing AI demand.

  Comparison

  CoWoS supports up to approximately 9.5 times reticle size, while Intel EMIB can support more than 12 times reticle size if execution proceeds smoothly.

  Risks

  Capacity expansion in CoWoS, advanced nodes, and supporting materials may lag demand, or customer capital expenditure may slow.

• Alchip

  Benefits from AWS Trainium custom AI chips and is included among the key AI names.

  Strengths

  Trainium revenue is expected to rise from US$1.8bn in 2026 to US$8bn in 2028.

  Weaknesses

  Trainium revenue is expected to account for 82% of total revenue in 2028, implying high customer and project concentration.

  Comparison

  Its growth primarily comes from AWS Trainium, while MediaTek's main incremental growth comes from Google TPU.

  Risks

  Trainium4 shipments, ASP recognition, or customer deployment progress could fall below assumptions.

• KYEC

  Benefits from chip-testing demand generated by growth in AI GPUs, TPUs, and CPUs and is included among the key AI names.

  Strengths

  Demand sources span general-purpose GPUs, custom ASICs, and server CPUs.

  Comparison

  Compared with a single chip design company, its revenue exposure lies in the testing segment required by multiple types of AI chips.

  Risks

  AI chip shipments may slow, or testing demand may materialize later than expected.

• AP Memory

  Identified as the Top Pick for memory and featured in the discussion of Intel EMIB-related advanced packaging technology.

  Strengths

  Has exposure to both memory and advanced-packaging-related technologies.

  Comparison

  The report includes it in the key memory list alongside Macronix, Nanya Tech, Winbond, and GigaDevice.

  Risks

  Improvement in memory supply-demand conditions or realization of advanced packaging projects may fall short of expectations.

• Hygon

  Explicitly rated OW and benefits from the expansion of China's CPU market and integrated CPU+GPU computing platforms.

  Strengths

  The report expects its share of China's server CPU market to reach 18% in 2028.

  Weaknesses

  Growth depends on China's server market and domestic computing infrastructure development.

  Comparison

  The report positions Hygon as an integrated CPU+GPU platform, while other domestic vendors focus more heavily on AI accelerators.

  Risks

  Constraints related to China's chip capacity, regulation, and the pace of computing infrastructure development.

• Cambricon、MetaX、Iluvatar

  All are associated with the themes of Chinese AI accelerators and domestic computing expansion; Cambricon and Iluvatar are included in the key China AI/semiconductor/WFE list.

  Strengths

  Domestic chips have lower prices, and the report believes they offer competitive total cost of ownership and cost performance.

  Weaknesses

  Advanced-chip capacity remains one of the constraints on AI growth in China.

  Comparison

  The report compares Cambricon, MetaX, and Iluvatar; MetaX is rated EW.

  Risks

  Risks related to chip capacity, ecosystem maturity, regulation, and project execution.

• OmniVision、Phison、MetaX、Realtek、GlobalWafers

  The report rates these companies EW.

  Weaknesses

  Their directional ratings are more neutral than those of the report's Top Picks and key beneficiaries.

  Comparison

  All are positioned as EW.

  Risks

  Company-specific risks are not fully elaborated in the provided research text.

• WIN Semi、Silergy、ASMedia

  The report rates these companies UW.

  Weaknesses

  The report assigns them an underweight directional positioning.

  Comparison

  Their ratings are more cautious than the Attractive industry view and the positioning of key AI names.

  Risks

  Company-specific risks are not fully elaborated in the provided research text.

Key data

• 2030 Global Semiconductor Market SizeUS$1.5tnThe report expects AI semiconductors to contribute approximately half.
• 2030 Global AI Semiconductor TAMApproximately US$753bnThe report's long-term market-size forecast.
• 2026 Cloud AI Semiconductor TAM Bull-Case ScenarioUS$485bnDerived from supply-chain data.
• Capital Expenditure Growth of the Four Largest CSPs87% YoY growth in the second quarter of 2026Covering Amazon, Google, Microsoft, and Meta.
• CSP Capital Expenditure-to-EBITDA RatioMore than 70%Indicates a high level of AI infrastructure investment intensity.
• 2027 Global Cloud Capital ExpenditureNearly US$1.4tnCovering the 14 largest listed global CSPs and excluding sovereign AI.
• 2027 AI Compute Wafer ConsumptionMore than US$46bnNVIDIA is expected to account for the majority.
• 2027 HBM ConsumptionUp to 48bn GbNVIDIA continues to consume most of the supply.
• TSMC N2 Capacity Growth70% CAGR from 2026 to 2028Based on the capacity roadmap provided by TSMC.
• TSMC AI Semiconductor Revenue ShareMore than 30% in 2026Share of the company's projected revenue.
• Alchip Trainium RevenueUS$8bn in 2028Approximately 82% of its projected US$9.8bn total revenue.
• MediaTek TPU RevenueUS$13.5bn in 2027; US$43.5bn in 2028; US$70bn in 2029Even under a customer-owned-tool model, the report still expects MediaTek to undertake part of the packaging and I/O die work.
• Agentic CPU TAM Growth251% CAGR from F26 to F30Morgan Stanley's top-down model; the bull-case scenario is US$238bn.
• China AI Chip TAMUS$91bn in 2030Driven by demand from cloud service providers, telecom operators, sovereign entities, and state-owned enterprises.
• China CPU TAMUS$42bn in 2030Hygon's share of China's server CPU market is expected to reach 18% in 2028.
• Atlas 950 SuperPod Scale1,024 Ascend 950DT NPUsShowcased at WAIC in 2026, supporting up to 256TB of pooled memory.
• PUF TAMApproximately US$310mn in 2030A hardware root-of-trust opportunity within the quantum hardware security market.

Impact & implications

The report believes AI capital expenditure will continue to concentrate revenue and capacity in advanced nodes, CoWoS, HBM, custom ASICs, server CPUs, and testing, with companies such as TSMC, MediaTek, Alchip, and KYEC benefiting more directly due to clear project or capacity exposure. Meanwhile, non-AI chips could be affected by resource crowding-out and cost inflation. In China, low-cost inference, improving domestic chip supply, and SuperPod clustering are expanding domestic computing opportunities; PQC migration is creating new demand for PUF and hardware security IP toward 2030.

Risks

• AI infrastructure budgets could limit continued rapid growth in capital expenditure.
• Data-center energy supply in the United States could constrain AI computing deployments.
• Advanced-chip capacity in China could limit the realization of domestic AI demand.
• Regulatory changes could affect AI chip investment, transactions, or supply-chain deployment.
• Rising wafer, assembly and testing, and memory costs could suppress technology product demand through price elasticity and squeeze chip design companies' margins in 2026.
• Prioritizing supply-chain resources for AI semiconductors could create shortages of resources such as T-Glass and memory and crowd out non-AI products.
• ABF substrate supply could constrain the volume ramp of AI chips such as Google TPU.
• Excluding memory and NVIDIA AI GPUs, non-AI semiconductor revenue growth is expected to decline in 2026.

What to watch

• Track global CSP capital expenditure growth, the capital expenditure-to-EBITDA ratio, and the actual implementation of announced power deployments.
• Monitor TSMC's progress in expanding CoWoS capacity to 200kwpm by 2027 and changes in the N2, N3, and N5 capacity mix.
• Track GB200/300 NVL72 rack shipments, AI compute wafer consumption, and HBM supply and demand in 2027.
• Monitor Trainium4, subsequent Google TPU generations, and the pace of revenue realization for MediaTek and Alchip projects.
• Track the duration of NAND, NOR Flash, and DDR4 shortages, as well as changes in spot and contract prices.
• Monitor China AI GPU demand, token usage, domestic chip capacity, and SuperPod deployment progress.
• Observe whether regional PQC migration roadmaps can proceed as planned before 2030.